Beyond technical efficacy: challenges and critical concerns of large language model’s impact on medical education in China: a systematic review

Introduction

Mainstream Large language model (LLM) platforms represented by ChatGPT-4o, Gork-3, SuperGPQA and DeepSeek-V3 had demonstrated mighty capabilities in medical text comprehension, clinical reasoning simulation, adaptive content generation, etc. [1], 2]. Globally, LLM was applicable to the specific context of medical education to enhance teaching effect. Institutions such as Harvard Medical School had integrated ChatGPT into case-based learning (CBL) and flipped classroom pedagogy, achieving a notable improvement in diagnostic accuracy among students [3]; meanwhile Google’s Med-PaLM two had attained “expert-level” performance (85 %) on U.S. medical licensing exams (USMLE), compared to its predecessor Med-PaLM, which attained only 60 % accuracy on similar medical licensing questions in USMLE [4]. In China, an original study assessed the performance of different LLM on a question bank designed for Chinese National Medical Licensing Examination in 2020 and 2021 (2020 and 2021 NMLE), GPT-4o achieved accuracy rates of 84.2 and 88.2 %, demonstrating a significant higher overall accuracy than GPT-4 and GPT-3.5 (p<0.001) [5]; another study reviewed the performance of LLMs from Chinese companies/western countries to answer questions in the National Traditional Chinese Medicine Licensing Examination (TCMLE) and evaluated the model’s explainability in answering traditional Chinese medicine (TCM)-related questions to determine its suitability as a TCM learning tool [6]. Currently a promising alternative is to leverage the capabilities of LLM to assist the communication in medical center reception sites in outpatient and preclinical education. From interactive question answering or academic writing refinement to medical education and training, in essence, LLM is crafted to emulate human-like comprehension and capabilities in text generation encompassing diverse content formats, in line with the development of medical teaching resources in China both in terms of functions and application scenarios [7]. These innovations highlight LLM’s potential to address systemic challenges like clinic bedside teaching shortages and resource disparities in preclinical to clinical decision-making process [8]. It is also an encouraging performance to receive attentions from the public for wider attempts of LLM in specific context of medical classroom teaching pedagogies (such as flipped classroom, team-based learning [TBL], problem-based learning [PBL], and CBL), to further verify the quality of medical innovation teaching [9].

The Stanford University Human-Centered Artificial Intelligence Index 2025 highlighted a global surge in public LLM optimism yet regional or professional division remained [10]. The report particularly expands its coverage of the role of LLM in the fields of science and medicine. LLM is transforming from a “data analysis tool” to a “research partner,” capable of participating in hypothesis generation, experimental design, interpretation of complex systems, and even conducting creative research directly involving medical education. The related Nobel Prizes and Turing Awards also confirmed LLM’s contribution to fundamental science and healthcare business [11]. Still, technical limitations, ethical and critical concerns still exist, such as model “hallucinations,” [12] lack of common reasoning, poor explainability, and multi-faceted ethical challenges in medical teaching, which are all key obstacles to the application of LLM in high-risk, high-reliability medical educational scenarios [13]. To better utilize LLM in further medical classroom teaching pedagogies of China, assessing the true capabilities and limitations of LLM itself is also a continuously developing research field [14]. Medical education in China faces concurrent demands for digital education and humanistic quality education [15]. This study aims to conduct a comprehensive review of scientific literature examining LLM’s applications in educational contexts, with particular emphasis on their implementation within China’s medical education system, which can ultimately advance the development of LLM systems to synergize advanced computational capabilities with localized educational corpus in medical training [16].

Methods

This systematic review employed the Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for scoping reviews (PRISMA-ScR) [17] framework to investigate: (1) LLMs’ implementation efficacy in China, (2) technical works and opportunities, (3) limitations and ethical compliance from campus to society, and (4) cultural collision and future orientations [18]. Both English and Chinese language databases were comprehensively searched in accordance with the query words for publications from PubMed, Web of Science (two English databases), CNKI, and Wanfang (two Chinese databases). The initial search period was limited from 1st Nov 2022 (related the release of ChatGPT 3.5 in November 2022) to 30th Apr 2025. The medical subject headings terms and search strategy keywords had been attached in Appendix file (Supplementary Appendix 1). The systematic literature review commenced with comprehensive database searches followed by duplicate removal using EndNote (Clarivate Analytics, version X9) as reference management software. The exclusion-inclusion decision protocol was documented in a PRISMA-compliant flow diagram (Figure 1) for transparency [19].

Figure 1:

The flowchart of the review search and scoping selection process. Abbreviations: WOS, web of science; CNKI, China national knowledge infrastructure; LLM, large language model.

The initial database search retrieved 876 publications (Supplementary Appendix 1). By additional manual search, 39 additional publications were identified and included in the selection process (the selection process was attached in Supplementary Appendix 2). Materials were reviewed and grouped into categories mainly from English publication databases, Chinese publication databases, to authorized academic conference documents. After removing duplicates and records lacking thematic relevance, 13 publications were retained for screening. In the updated literature search, three additional publications were identified. After reviewing the titles and abstracts, one additional article underwent full-text evaluation and was included in the cohort. Consequently, A total of 14 selected publications with high representativeness or topic relevance were used for subsequent information collation and discussion (the basic information of involved 14 literatures had been attached in Table 1Table 1:), and the basic information and core themes were summarized according to their topics. The flowchart of the search and selection process was attached in Figure 1.

Results

A total of 14 full-text journal articles, reviews, and comments with high average quotation rate or regional educational characteristics were extracted from 915 retrieved studies. Study types included original research articles (n=3, 21.4 %), reviews articles (n=5, 35.7 %), and perspective/commentary/viewpoint articles (n=5, 35.7 %). Additionally, one academic report was included (n=1, 7.1 %). The basic information had been attached by category in Table 1 with corresponding specific focus themes been briefly illustrated. All 14 included studies identified the inevitable positive effects of LLMs on digital health care educational module, demonstrated the pluralistic perspective from medical education details [28].

Discussion

This systematic review illuminates the intricate dynamics of integrating LLMs into China’s medical education system, highlighting both their transformative potential and the persistent challenges that shape their adoption [25]. Drawing on evidence from the 14 studies included in this review, this discussion interprets the findings in the context of the existing literature, addresses their implications, and offers recommendations for educators, learners, and developers to standardize and optimize LLM use in medical education.

As discussed by Sandmann et al. [41], medical faculty and students have shown significant progression: moving beyond simply lacking LLM-based support in clinical decision-making, they are now actively engaged in identifying issues, training LLMs to enhance clinical abilities, and comparing performance details. There was a significant increase in awareness of open-source LLM related applications or knowledges, not only among faculty or academic researchers in all fields, but among the public. Similarly, as Jiang et al. [7] pointed out, clinical medicine professors had already recognized potential issues with improper LLM usage, such as academic dishonesty, discrepancies between generated results and clinical reality, and LLMs’ inability to replace actual hospital workflows or specific steps in research experiments. The growing recognition and discussion of these concerns and insights represent an encouraging development. Compared to the previous, there was a subtle shift in attitudes toward LLM among scholars and the public in China between 2024 and 2025 [25]. The 2025 AI Index Annual Report provided detailed data on economic investments, showing a global trend of “investing in more advanced and powerful LLMs.” This reflects complex critical thinking, tempered by prevailing optimism. The progress of LLMs was tangible, benefiting many medical faculty and students in areas such as extensive research, teaching, machine recognition in clinical systems, article reviews, and text editing. This aligned well with the essential needs of medical education reform in China, and the standardization of related educational courses seemed foreseeable. In medical curricula, LLMs were primarily used to test clinical cases in specific specialties, training clinical logical thinking through these cases [26], 27]. However, LLM systems still faced delays in learning, especially when dealing with newly published articles. For instance, when queried about a recently published Nature Methods paper on unsupervised domain adaptation for sequencing (UDA-seq) [45], LLMs still provided responses focused on outdated information regarding spatial transcriptomics or wrongly associated UDA-seq with bulk RNA sequencing or single-cell RNA sequencing (scRNA-seq) technologies, which was irrelevant to the topic. Extensive training was crucial, and we recommend establishing specialized medical databases to enhance precision. Balancing technical accuracy with medical humanities is crucial for optimal overall performance. In the classroom, China’s teaching methods such asCBL and flipped classrooms were closely linked with LLMs [45]. These methods aligned well with the demands of China’s National Health Commission and Ministry of Education to develop students, making the process of cultivating excellent medical students into physician-scientists in parallel with the direction of LLM development and integration in medical education.

From the students’ perspective, excessive reliance on LLMs undermined critical and clinical thinking, with practical skills potentially being neglected. This could hinder students’ ability to make independent, informed decisions in clinical practice. Moral dilemmas also stem from potential misalignments between LLM decision-making logic and established medical ethical standards, which could have affected the teaching of essential ethical principles to future healthcare professionals. The lack of humanistic care in current LLM tools was a critical gap, as these models often failed to convey emotional aspects and empathy. Training LLMs to incorporate humanistic elements could help students connect more deeply with patients. According to The 2025 AI Index Annual Report, educators, students, and the public all showed higher enthusiasm for the development and performance of LLMs, with fewer concerns compared to previous years. Addressing these ethical issues is essential to harness the benefits of LLMs in medical education while preserving fairness, accountability, and empathy, ultimately promoting the integration of “LLM clinical education” and establishing a robust oversight framework to ensure safe implementation in the medical field [27].

The long-term integration of LLM into medical education faces significant challenges. Despite its potential, Cooper et al. [14] conveyed judgements that LLM is still in the early stages of development, both technically and practically. While LLMs can enhance learning, they struggle with complex clinical scenarios, particularly in specialized fields that require nuanced, context-specific expertise. For instance, LLMs might encounter with difficulty interpreting culturally embedded knowledge such as pulse diagnosis in traditional Chinese medicine. These limitations highlight the need for ongoing testing and model refinement. Additionally, the legal framework for LLM in education remains underdeveloped, lacking clear accreditation, licensing, and oversight mechanisms, raising concerns about data privacy, algorithmic bias, and accountability. The rapid evolution of medical knowledge further complicates LLM integration, necessitating continuous updates and maintenance, which strains resources. Furthermore, a shortage of interdisciplinary talent capable of bridging AI development and medical pedagogy, coupled with the high cost of training such experts, compounds these challenges [46]. Integrating LLMs into already crowded medical curricula requires creative solutions to optimize teaching schedules without compromising quality. Despite these challenges, the potential of LLMs in medical education remains promising, though their practical application will require sustained collaboration across technology, education, and policy sectors [47].

Returning to the initial question, it is evident that standardized workflows or nationally scalable models (e.g., online open curricula) for integrating LLMs into medical education remain underdeveloped, particularly at the undergraduate level. In this paper, we compared LLM performance across three standardized medical examinations (the USMLE, China’s NMLE and TCMLE). They theoretically revealed discrepancies between LLMs without specific medical training and those fine-tuned on medical dataset, reflecting their respective mastery of core medical knowledge. During undergraduate self-directed learning, LLMs are primarily employed for manuscript drafting and refinement, exam preparation such as generating virtual patient cases or discipline-specific question sets, literature analysis, extracting structured outlines and key information from extensive text or image data, and reducing workload, especially for repetitive tasks. More specialized research and clinical task (including electronic health record management, pathology image interpretation, R script debugging, and developing models for specific clinical question) are currently primarily undertaken by graduate students within China’s mentor-based system. We propose implementing a general education course (online, offline, or hybrid) for junior undergraduates on effectively leveraging LLMs for both in-class and extracurricular learning. This approach is not only feasible but holds significant potential benefits. Simultaneously, it raises critical concerns regarding academic integrity, legal and regulatory compliance, and data privacy, necessitating collaboration among China’s relevant regulatory bodies and domain experts. Initially, undergraduates can utilize LLMs within CBL and flipped classroom to address specific clinical or research questions given at class. Subsequently, they can expand their explorations into review articles or innovation project proposals. Even students encountering research or clinical problems for the first time can gain insights by observing the problem-solving logic of LLMs, thereby enriching their own learning strategies [48]. To foster the LLM-assisted ecosystem, further collaboration among multiple universities or medical centers is essential. How to utilize their clinical data and teaching evaluable materials, and how to develop standardized protocols or publicly accessible online course modules, limitations or orientations mentioned above require adequate resources and scientific experimentation.

This study’s limitations include linguistic bias toward Chinese/English publications and urban-centric sampling. Limitations also existed in sample selection, inferences and conclusions, which might not be consistent with the reader’s affiliation. Future investigations should incorporate minority language resources and non-textual media such as clinical simulation recordings [49]. As China navigates these challenges, its experience offers unique insights into global efforts to harmonize technological progress with cultural preservation in medical education – a balance demanding continued innovation, rigorous oversight, and cross-disciplinary collaboration. China’s experience with LLM in medical education provides a unique lens on global efforts to integrate AI while preserving cultural and educational values. The findings align with international trends. By addressing these through interdisciplinary collaboration, technological iteration, and ethical oversight, China can contribute to a global model of AI-driven education that balances innovation with tradition.

Conclusions

The application of generative LLMs in China’s medical education demonstrates parallel trends of technological advancement and scenario expansion. Despite significant achievements in personalized teaching, case generation, and other fields, core challenges such as data security and assessment efficacy still require addressing through interdisciplinary collaboration, robust ethical standards, and continuous technological iteration. Future research should focus on building an educational ecosystem of “human-machine collaborative” educational ecosystem to drive the transformation of medical education towards better intelligence and precision.